4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2001, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
32 #define DEBUG_SUBSYSTEM S_MDC
34 #include <linux/init.h>
35 #include <linux/kthread.h>
36 #include <linux/module.h>
37 #include <linux/pagemap.h>
38 #include <linux/user_namespace.h>
39 #include <linux/utsname.h>
40 #include <linux/delay.h>
41 #include <linux/uidgid.h>
42 #include <linux/device.h>
43 #include <linux/xarray.h>
45 #include <lustre_errno.h>
47 #include <cl_object.h>
48 #include <llog_swab.h>
49 #include <lprocfs_status.h>
50 #include <lustre_acl.h>
51 #include <lustre_compat.h>
52 #include <lustre_fid.h>
53 #include <uapi/linux/lustre/lustre_ioctl.h>
54 #include <lustre_kernelcomm.h>
55 #include <lustre_lmv.h>
56 #include <lustre_log.h>
57 #include <lustre_swab.h>
58 #include <obd_class.h>
59 #include <lustre_osc.h>
61 #include "mdc_internal.h"
63 #define REQUEST_MINOR 244
65 static int mdc_cleanup(struct obd_device *obd);
67 static inline int mdc_queue_wait(struct ptlrpc_request *req)
69 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
72 /* obd_get_request_slot() ensures that this client has no more
73 * than cl_max_rpcs_in_flight RPCs simultaneously inf light
75 rc = obd_get_request_slot(cli);
79 rc = ptlrpc_queue_wait(req);
80 obd_put_request_slot(cli);
86 * Send MDS_GET_ROOT RPC to fetch root FID.
88 * If \a fileset is not NULL it should contain a subdirectory off
89 * the ROOT/ directory to be mounted on the client. Return the FID
90 * of the subdirectory to the client to mount onto its mountpoint.
92 * \param[in] imp MDC import
93 * \param[in] fileset fileset name, which could be NULL
94 * \param[out] rootfid root FID of this mountpoint
95 * \param[out] pc root capa will be unpacked and saved in this pointer
97 * \retval 0 on success, negative errno on failure
99 static int mdc_get_root(struct obd_export *exp, const char *fileset,
100 struct lu_fid *rootfid)
102 struct ptlrpc_request *req;
103 struct mdt_body *body;
108 if (fileset && !(exp_connect_flags(exp) & OBD_CONNECT_SUBTREE))
111 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
117 req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
118 strlen(fileset) + 1);
119 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GET_ROOT);
121 ptlrpc_request_free(req);
124 mdc_pack_body(&req->rq_pill, NULL, 0, 0, -1, 0);
125 if (fileset != NULL) {
126 char *name = req_capsule_client_get(&req->rq_pill, &RMF_NAME);
128 memcpy(name, fileset, strlen(fileset));
130 lustre_msg_add_flags(req->rq_reqmsg, LUSTRE_IMP_FULL);
131 req->rq_send_state = LUSTRE_IMP_FULL;
133 ptlrpc_request_set_replen(req);
135 rc = ptlrpc_queue_wait(req);
139 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
141 GOTO(out, rc = -EPROTO);
143 *rootfid = body->mbo_fid1;
144 CDEBUG(D_NET, "root fid="DFID", last_committed=%llu\n",
145 PFID(rootfid), lustre_msg_get_last_committed(req->rq_repmsg));
148 ptlrpc_req_finished(req);
154 * This function now is known to always saying that it will receive 4 buffers
155 * from server. Even for cases when acl_size and md_size is zero, RPC header
156 * will contain 4 fields and RPC itself will contain zero size fields. This is
157 * because mdt_getattr*() _always_ returns 4 fields, but if acl is not needed
158 * and thus zero, it shrinks it, making zero size. The same story about
159 * md_size. And this is course of problem when client waits for smaller number
160 * of fields. This issue will be fixed later when client gets aware of RPC
163 static int mdc_getattr_common(struct obd_export *exp,
164 struct ptlrpc_request *req,
165 struct md_op_data *op_data)
167 struct req_capsule *pill = &req->rq_pill;
168 struct mdt_body *body;
173 /* Request message already built. */
174 rc = ptlrpc_queue_wait(req);
178 /* sanity check for the reply */
179 body = req_capsule_server_get(pill, &RMF_MDT_BODY);
183 CDEBUG(D_NET, "mode: %o\n", body->mbo_mode);
185 mdc_update_max_ea_from_body(exp, body);
186 if (body->mbo_eadatasize != 0) {
187 eadata = req_capsule_server_sized_get(pill, &RMF_MDT_MD,
188 body->mbo_eadatasize);
193 /* If encryption context was returned by MDT, put it in op_data
194 * so that caller can set it on inode and save an extra getxattr.
196 if (op_data && op_data->op_valid & OBD_MD_ENCCTX &&
197 body->mbo_valid & OBD_MD_ENCCTX) {
198 op_data->op_file_encctx =
199 req_capsule_server_get(pill, &RMF_FILE_ENCCTX);
200 op_data->op_file_encctx_size =
201 req_capsule_get_size(pill, &RMF_FILE_ENCCTX,
208 static void mdc_reset_acl_req(struct ptlrpc_request *req)
210 spin_lock(&req->rq_early_free_lock);
211 sptlrpc_cli_free_repbuf(req);
212 req->rq_repbuf = NULL;
213 req->rq_repbuf_len = 0;
214 req->rq_repdata = NULL;
215 req->rq_reqdata_len = 0;
216 spin_unlock(&req->rq_early_free_lock);
219 static int mdc_getattr(struct obd_export *exp, struct md_op_data *op_data,
220 struct ptlrpc_request **request)
222 struct ptlrpc_request *req;
223 struct obd_device *obd = class_exp2obd(exp);
224 struct obd_import *imp = class_exp2cliimp(exp);
225 __u32 acl_bufsize = LUSTRE_POSIX_ACL_MAX_SIZE_OLD;
229 /* Single MDS without an LMV case */
230 if (op_data->op_flags & MF_GET_MDT_IDX) {
236 req = ptlrpc_request_alloc(imp, &RQF_MDS_GETATTR);
240 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GETATTR);
242 ptlrpc_request_free(req);
246 /* LU-15245: avoid deadlock with modifying RPCs on MDS_REQUEST_PORTAL */
247 req->rq_request_portal = MDS_READPAGE_PORTAL;
250 mdc_pack_body(&req->rq_pill, &op_data->op_fid1, op_data->op_valid,
251 op_data->op_mode, -1, 0);
252 req_capsule_set_size(&req->rq_pill, &RMF_ACL, RCL_SERVER, acl_bufsize);
253 req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
255 if (exp_connect_encrypt(exp) && op_data->op_valid & OBD_MD_ENCCTX)
256 req_capsule_set_size(&req->rq_pill, &RMF_FILE_ENCCTX,
258 obd->u.cli.cl_max_mds_easize);
260 req_capsule_set_size(&req->rq_pill, &RMF_FILE_ENCCTX,
262 ptlrpc_request_set_replen(req);
264 rc = mdc_getattr_common(exp, req, op_data);
267 acl_bufsize = min_t(__u32,
268 imp->imp_connect_data.ocd_max_easize,
270 mdc_reset_acl_req(req);
274 ptlrpc_req_finished(req);
282 static int mdc_getattr_name(struct obd_export *exp, struct md_op_data *op_data,
283 struct ptlrpc_request **request)
285 struct ptlrpc_request *req;
286 struct obd_import *imp = class_exp2cliimp(exp);
287 __u32 acl_bufsize = LUSTRE_POSIX_ACL_MAX_SIZE_OLD;
292 req = ptlrpc_request_alloc(imp, &RQF_MDS_GETATTR_NAME);
296 req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
297 op_data->op_namelen + 1);
299 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GETATTR_NAME);
301 ptlrpc_request_free(req);
305 if (op_data->op_name) {
306 char *name = req_capsule_client_get(&req->rq_pill, &RMF_NAME);
307 LASSERT(strnlen(op_data->op_name, op_data->op_namelen) ==
308 op_data->op_namelen);
309 memcpy(name, op_data->op_name, op_data->op_namelen);
313 mdc_pack_body(&req->rq_pill, &op_data->op_fid1, op_data->op_valid,
314 op_data->op_mode, op_data->op_suppgids[0], 0);
315 req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
317 req_capsule_set_size(&req->rq_pill, &RMF_ACL, RCL_SERVER, acl_bufsize);
318 req_capsule_set_size(&req->rq_pill, &RMF_FILE_ENCCTX, RCL_SERVER, 0);
319 ptlrpc_request_set_replen(req);
320 if (op_data->op_bias & MDS_FID_OP) {
321 struct mdt_body *b = req_capsule_client_get(&req->rq_pill,
325 b->mbo_valid |= OBD_MD_NAMEHASH;
326 b->mbo_fid2 = op_data->op_fid2;
330 rc = mdc_getattr_common(exp, req, NULL);
333 acl_bufsize = min_t(__u32,
334 imp->imp_connect_data.ocd_max_easize,
336 mdc_reset_acl_req(req);
340 ptlrpc_req_finished(req);
348 static int mdc_xattr_common(struct obd_export *exp,const struct req_format *fmt,
349 const struct lu_fid *fid, int opcode, u64 valid,
350 const char *xattr_name, const char *input,
351 int input_size, int output_size, int flags,
352 __u32 suppgid, struct ptlrpc_request **request)
354 struct ptlrpc_request *req;
355 int xattr_namelen = 0;
361 req = ptlrpc_request_alloc(class_exp2cliimp(exp), fmt);
366 xattr_namelen = strlen(xattr_name) + 1;
367 req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
372 req_capsule_set_size(&req->rq_pill, &RMF_EADATA, RCL_CLIENT,
375 /* get SELinux policy info if any */
376 rc = sptlrpc_get_sepol(req);
378 ptlrpc_request_free(req);
381 req_capsule_set_size(&req->rq_pill, &RMF_SELINUX_POL, RCL_CLIENT,
382 strlen(req->rq_sepol) ?
383 strlen(req->rq_sepol) + 1 : 0);
385 /* Flush local XATTR locks to get rid of a possible cancel RPC */
386 if (opcode == MDS_REINT && fid_is_sane(fid) &&
387 exp->exp_connect_data.ocd_ibits_known & MDS_INODELOCK_XATTR) {
391 /* Without that packing would fail */
393 req_capsule_set_size(&req->rq_pill, &RMF_EADATA,
396 count = mdc_resource_get_unused(exp, fid,
398 MDS_INODELOCK_XATTR);
400 rc = mdc_prep_elc_req(exp, req, MDS_REINT, &cancels, count);
402 ptlrpc_request_free(req);
406 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, opcode);
408 ptlrpc_request_free(req);
413 if (opcode == MDS_REINT) {
414 struct mdt_rec_setxattr *rec;
416 BUILD_BUG_ON(sizeof(struct mdt_rec_setxattr) !=
417 sizeof(struct mdt_rec_reint));
418 rec = req_capsule_client_get(&req->rq_pill, &RMF_REC_REINT);
419 rec->sx_opcode = REINT_SETXATTR;
420 rec->sx_fsuid = from_kuid(&init_user_ns, current_fsuid());
421 rec->sx_fsgid = from_kgid(&init_user_ns, current_fsgid());
422 rec->sx_cap = current_cap().cap[0];
423 rec->sx_suppgid1 = suppgid;
424 rec->sx_suppgid2 = -1;
426 rec->sx_valid = valid | OBD_MD_FLCTIME;
427 rec->sx_time = ktime_get_real_seconds();
428 rec->sx_size = output_size;
429 rec->sx_flags = flags;
431 mdc_pack_body(&req->rq_pill, fid, valid, output_size,
433 /* Avoid deadlock with modifying RPCs on MDS_REQUEST_PORTAL.
436 req->rq_request_portal = MDS_READPAGE_PORTAL;
440 tmp = req_capsule_client_get(&req->rq_pill, &RMF_NAME);
441 memcpy(tmp, xattr_name, xattr_namelen);
444 tmp = req_capsule_client_get(&req->rq_pill, &RMF_EADATA);
445 memcpy(tmp, input, input_size);
448 mdc_file_sepol_pack(&req->rq_pill);
450 if (req_capsule_has_field(&req->rq_pill, &RMF_EADATA, RCL_SERVER))
451 req_capsule_set_size(&req->rq_pill, &RMF_EADATA,
452 RCL_SERVER, output_size);
453 ptlrpc_request_set_replen(req);
456 if (opcode == MDS_REINT)
457 ptlrpc_get_mod_rpc_slot(req);
459 rc = ptlrpc_queue_wait(req);
461 if (opcode == MDS_REINT)
462 ptlrpc_put_mod_rpc_slot(req);
465 ptlrpc_req_finished(req);
471 static int mdc_setxattr(struct obd_export *exp, const struct lu_fid *fid,
472 u64 obd_md_valid, const char *name,
473 const void *value, size_t value_size,
474 unsigned int xattr_flags, u32 suppgid,
475 struct ptlrpc_request **req)
477 LASSERT(obd_md_valid == OBD_MD_FLXATTR ||
478 obd_md_valid == OBD_MD_FLXATTRRM);
480 return mdc_xattr_common(exp, &RQF_MDS_REINT_SETXATTR,
481 fid, MDS_REINT, obd_md_valid, name,
482 value, value_size, 0, xattr_flags, suppgid,
486 static int mdc_getxattr(struct obd_export *exp, const struct lu_fid *fid,
487 u64 obd_md_valid, const char *name, size_t buf_size,
488 struct ptlrpc_request **req)
490 struct mdt_body *body;
493 LASSERT(obd_md_valid == OBD_MD_FLXATTR ||
494 obd_md_valid == OBD_MD_FLXATTRLS);
496 /* Message below is checked in sanity-selinux test_20d
497 * and sanity-sec test_49
499 CDEBUG(D_INFO, "%s: get xattr '%s' for "DFID"\n",
500 exp->exp_obd->obd_name, name, PFID(fid));
501 rc = mdc_xattr_common(exp, &RQF_MDS_GETXATTR, fid, MDS_GETXATTR,
502 obd_md_valid, name, NULL, 0, buf_size, 0, -1,
507 body = req_capsule_server_get(&(*req)->rq_pill, &RMF_MDT_BODY);
509 GOTO(out, rc = -EPROTO);
511 /* only detect the xattr size */
513 /* LU-11109: Older MDTs do not distinguish
514 * between nonexistent xattrs and zero length
515 * values in this case. Newer MDTs will return
516 * -ENODATA or set OBD_MD_FLXATTR. */
517 GOTO(out, rc = body->mbo_eadatasize);
520 if (body->mbo_eadatasize == 0) {
521 /* LU-11109: Newer MDTs set OBD_MD_FLXATTR on
522 * success so that we can distinguish between
523 * zero length value and nonexistent xattr.
525 * If OBD_MD_FLXATTR is not set then we keep
526 * the old behavior and return -ENODATA for
527 * getxattr() when mbo_eadatasize is 0. But
528 * -ENODATA only makes sense for getxattr()
529 * and not for listxattr(). */
530 if (body->mbo_valid & OBD_MD_FLXATTR)
532 else if (obd_md_valid == OBD_MD_FLXATTR)
533 GOTO(out, rc = -ENODATA);
538 GOTO(out, rc = body->mbo_eadatasize);
541 ptlrpc_req_finished(*req);
548 static int mdc_get_lustre_md(struct obd_export *exp, struct req_capsule *pill,
549 struct obd_export *dt_exp,
550 struct obd_export *md_exp,
551 struct lustre_md *md)
557 memset(md, 0, sizeof(*md));
559 md->body = req_capsule_server_get(pill, &RMF_MDT_BODY);
560 LASSERT(md->body != NULL);
562 if (md->body->mbo_valid & OBD_MD_FLEASIZE) {
563 if (!S_ISREG(md->body->mbo_mode)) {
564 CDEBUG(D_INFO, "OBD_MD_FLEASIZE set, should be a "
565 "regular file, but is not\n");
566 GOTO(out, rc = -EPROTO);
569 if (md->body->mbo_eadatasize == 0) {
570 CDEBUG(D_INFO, "OBD_MD_FLEASIZE set, "
571 "but eadatasize 0\n");
572 GOTO(out, rc = -EPROTO);
575 md->layout.lb_len = md->body->mbo_eadatasize;
576 md->layout.lb_buf = req_capsule_server_sized_get(pill,
579 if (md->layout.lb_buf == NULL)
580 GOTO(out, rc = -EPROTO);
581 } else if (md->body->mbo_valid & OBD_MD_FLDIREA) {
582 const union lmv_mds_md *lmv;
585 if (!S_ISDIR(md->body->mbo_mode)) {
586 CDEBUG(D_INFO, "OBD_MD_FLDIREA set, should be a "
587 "directory, but is not\n");
588 GOTO(out, rc = -EPROTO);
591 if (md_exp->exp_obd->obd_type->typ_lu == &mdc_device_type) {
592 CERROR("%s: no LMV, upgrading from old version?\n",
593 md_exp->exp_obd->obd_name);
595 GOTO(out_acl, rc = 0);
598 if (md->body->mbo_valid & OBD_MD_MEA) {
599 lmv_size = md->body->mbo_eadatasize;
601 CDEBUG(D_INFO, "OBD_MD_FLDIREA is set, "
602 "but eadatasize 0\n");
606 lmv = req_capsule_server_sized_get(pill, &RMF_MDT_MD,
609 GOTO(out, rc = -EPROTO);
611 rc = md_unpackmd(md_exp, &md->lmv, lmv, lmv_size);
615 if (rc < (int)sizeof(*md->lmv)) {
616 struct lmv_foreign_md *lfm = md->lfm;
618 /* short (< sizeof(struct lmv_stripe_md))
621 if (lfm->lfm_magic != LMV_MAGIC_FOREIGN) {
623 "lmv size too small: %d < %d\n",
624 rc, (int)sizeof(*md->lmv));
625 GOTO(out, rc = -EPROTO);
630 /* since 2.12.58 intent_getattr fetches default LMV */
631 if (md->body->mbo_valid & OBD_MD_DEFAULT_MEA) {
632 lmv_size = sizeof(struct lmv_user_md);
633 lmv = req_capsule_server_sized_get(pill,
637 GOTO(out, rc = -EPROTO);
639 rc = md_unpackmd(md_exp, &md->default_lmv, lmv,
644 if (rc < (int)sizeof(*md->default_lmv)) {
646 "default lmv size too small: %d < %d\n",
647 rc, (int)sizeof(*md->default_lmv));
648 GOTO(out, rc = -EPROTO);
655 if (md->body->mbo_valid & OBD_MD_FLACL) {
656 /* for ACL, it's possible that FLACL is set but aclsize is zero.
657 * only when aclsize != 0 there's an actual segment for ACL
660 rc = mdc_unpack_acl(pill, md);
673 static int mdc_free_lustre_md(struct obd_export *exp, struct lustre_md *md)
679 void mdc_replay_open(struct ptlrpc_request *req)
681 struct md_open_data *mod = req->rq_cb_data;
682 struct ptlrpc_request *close_req;
683 struct obd_client_handle *och;
684 struct lustre_handle old_open_handle = { };
685 struct mdt_body *body;
686 struct ldlm_reply *rep;
690 DEBUG_REQ(D_ERROR, req,
691 "cannot properly replay without open data");
696 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
697 LASSERT(body != NULL);
699 rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
700 if (rep != NULL && rep->lock_policy_res2 != 0)
701 DEBUG_REQ(D_ERROR, req, "Open request replay failed with %ld ",
702 (long int)rep->lock_policy_res2);
704 spin_lock(&req->rq_lock);
706 if (och && och->och_open_handle.cookie)
707 req->rq_early_free_repbuf = 1;
709 req->rq_early_free_repbuf = 0;
710 spin_unlock(&req->rq_lock);
712 if (req->rq_early_free_repbuf) {
713 struct lustre_handle *file_open_handle;
715 LASSERT(och->och_magic == OBD_CLIENT_HANDLE_MAGIC);
717 file_open_handle = &och->och_open_handle;
718 CDEBUG(D_HA, "updating handle from %#llx to %#llx\n",
719 file_open_handle->cookie, body->mbo_open_handle.cookie);
720 old_open_handle = *file_open_handle;
721 *file_open_handle = body->mbo_open_handle;
724 close_req = mod->mod_close_req;
726 __u32 opc = lustre_msg_get_opc(close_req->rq_reqmsg);
727 struct mdt_ioepoch *epoch;
729 LASSERT(opc == MDS_CLOSE);
730 epoch = req_capsule_client_get(&close_req->rq_pill,
734 if (req->rq_early_free_repbuf)
735 LASSERT(old_open_handle.cookie ==
736 epoch->mio_open_handle.cookie);
738 DEBUG_REQ(D_HA, close_req, "updating close body with new fh");
739 epoch->mio_open_handle = body->mbo_open_handle;
744 void mdc_commit_open(struct ptlrpc_request *req)
746 struct md_open_data *mod = req->rq_cb_data;
751 * No need to touch md_open_data::mod_och, it holds a reference on
752 * \var mod and will zero references to each other, \var mod will be
753 * freed after that when md_open_data::mod_och will put the reference.
757 * Do not let open request to disappear as it still may be needed
758 * for close rpc to happen (it may happen on evict only, otherwise
759 * ptlrpc_request::rq_replay does not let mdc_commit_open() to be
760 * called), just mark this rpc as committed to distinguish these 2
761 * cases, see mdc_close() for details. The open request reference will
762 * be put along with freeing \var mod.
764 ptlrpc_request_addref(req);
765 spin_lock(&req->rq_lock);
766 req->rq_committed = 1;
767 spin_unlock(&req->rq_lock);
768 req->rq_cb_data = NULL;
772 int mdc_set_open_replay_data(struct obd_export *exp,
773 struct obd_client_handle *och,
774 struct lookup_intent *it)
776 struct md_open_data *mod;
777 struct mdt_rec_create *rec;
778 struct mdt_body *body;
779 struct ptlrpc_request *open_req = it->it_request;
780 struct obd_import *imp = open_req->rq_import;
783 if (!open_req->rq_replay)
786 rec = req_capsule_client_get(&open_req->rq_pill, &RMF_REC_REINT);
787 body = req_capsule_server_get(&open_req->rq_pill, &RMF_MDT_BODY);
788 LASSERT(rec != NULL);
789 /* Incoming message in my byte order (it's been swabbed). */
790 /* Outgoing messages always in my byte order. */
791 LASSERT(body != NULL);
793 /* Only if the import is replayable, we set replay_open data */
794 if (och && imp->imp_replayable) {
795 mod = obd_mod_alloc();
797 DEBUG_REQ(D_ERROR, open_req,
798 "cannot allocate md_open_data");
803 * Take a reference on \var mod, to be freed on mdc_close().
804 * It protects \var mod from being freed on eviction (commit
805 * callback is called despite rq_replay flag).
806 * Another reference for \var och.
811 spin_lock(&open_req->rq_lock);
814 mod->mod_is_create = it_disposition(it, DISP_OPEN_CREATE) ||
815 it_disposition(it, DISP_OPEN_STRIPE);
816 mod->mod_open_req = open_req;
817 open_req->rq_cb_data = mod;
818 open_req->rq_commit_cb = mdc_commit_open;
819 open_req->rq_early_free_repbuf = 1;
820 spin_unlock(&open_req->rq_lock);
823 rec->cr_fid2 = body->mbo_fid1;
824 rec->cr_open_handle_old = body->mbo_open_handle;
825 open_req->rq_replay_cb = mdc_replay_open;
826 if (!fid_is_sane(&body->mbo_fid1)) {
827 DEBUG_REQ(D_ERROR, open_req,
828 "saving replay request with insane FID " DFID,
829 PFID(&body->mbo_fid1));
833 DEBUG_REQ(D_RPCTRACE, open_req, "Set up open replay data");
837 static void mdc_free_open(struct md_open_data *mod)
841 if (mod->mod_is_create == 0 &&
842 imp_connect_disp_stripe(mod->mod_open_req->rq_import))
846 * No reason to asssert here if the open request has
847 * rq_replay == 1. It means that mdc_close failed, and
848 * close request wasn`t sent. It is not fatal to client.
849 * The worst thing is eviction if the client gets open lock
852 DEBUG_REQ(D_RPCTRACE, mod->mod_open_req,
853 "free open request, rq_replay=%d",
854 mod->mod_open_req->rq_replay);
856 ptlrpc_request_committed(mod->mod_open_req, committed);
857 if (mod->mod_close_req)
858 ptlrpc_request_committed(mod->mod_close_req, committed);
861 static int mdc_clear_open_replay_data(struct obd_export *exp,
862 struct obd_client_handle *och)
864 struct md_open_data *mod = och->och_mod;
868 * It is possible to not have \var mod in a case of eviction between
869 * lookup and ll_file_open().
874 LASSERT(mod != LP_POISON);
875 LASSERT(mod->mod_open_req != NULL);
877 spin_lock(&mod->mod_open_req->rq_lock);
879 mod->mod_och->och_open_handle.cookie = 0;
880 mod->mod_open_req->rq_early_free_repbuf = 0;
881 spin_unlock(&mod->mod_open_req->rq_lock);
891 static int mdc_close(struct obd_export *exp, struct md_op_data *op_data,
892 struct md_open_data *mod, struct ptlrpc_request **request)
894 struct obd_device *obd = class_exp2obd(exp);
895 struct ptlrpc_request *req;
896 struct req_format *req_fmt;
897 size_t u32_count = 0;
902 CDEBUG(D_INODE, "%s: "DFID" file closed with intent: %x\n",
903 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
906 if (op_data->op_bias & MDS_CLOSE_INTENT) {
907 req_fmt = &RQF_MDS_CLOSE_INTENT;
908 if (op_data->op_bias & MDS_HSM_RELEASE) {
909 /* allocate a FID for volatile file */
910 rc = mdc_fid_alloc(NULL, exp, &op_data->op_fid2,
913 CERROR("%s: "DFID" allocating FID: rc = %d\n",
914 obd->obd_name, PFID(&op_data->op_fid1),
916 /* save the errcode and proceed to close */
920 if (op_data->op_bias & MDS_CLOSE_RESYNC_DONE) {
921 size_t count = op_data->op_data_size / sizeof(__u32);
923 if (count > INLINE_RESYNC_ARRAY_SIZE)
927 req_fmt = &RQF_MDS_CLOSE;
931 if (OBD_FAIL_CHECK(OBD_FAIL_MDC_CLOSE))
934 req = ptlrpc_request_alloc(class_exp2cliimp(exp), req_fmt);
936 /* Ensure that this close's handle is fixed up during replay. */
937 if (likely(mod != NULL)) {
938 LASSERTF(mod->mod_open_req != NULL &&
939 mod->mod_open_req->rq_type != LI_POISON,
940 "POISONED open %p!\n", mod->mod_open_req);
942 mod->mod_close_req = req;
944 DEBUG_REQ(D_RPCTRACE, mod->mod_open_req, "matched open");
945 /* We no longer want to preserve this open for replay even
946 * though the open was committed. b=3632, b=3633 */
947 spin_lock(&mod->mod_open_req->rq_lock);
948 mod->mod_open_req->rq_replay = 0;
949 spin_unlock(&mod->mod_open_req->rq_lock);
951 CDEBUG(D_HA, "couldn't find open req; expecting close error\n");
955 * TODO: repeat close after errors
957 CWARN("%s: close of FID "DFID" failed, file reference will be "
958 "dropped when this client unmounts or is evicted\n",
959 obd->obd_name, PFID(&op_data->op_fid1));
960 GOTO(out, rc = -ENOMEM);
964 req_capsule_set_size(&req->rq_pill, &RMF_U32, RCL_CLIENT,
965 u32_count * sizeof(__u32));
967 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_CLOSE);
969 ptlrpc_request_free(req);
974 /* To avoid a livelock (bug 7034), we need to send CLOSE RPCs to a
975 * portal whose threads are not taking any DLM locks and are therefore
976 * always progressing */
977 req->rq_request_portal = MDS_READPAGE_PORTAL;
978 ptlrpc_at_set_req_timeout(req);
980 if (!obd->u.cli.cl_lsom_update ||
981 !(exp_connect_flags2(exp) & OBD_CONNECT2_LSOM))
982 op_data->op_xvalid &= ~(OP_XVALID_LAZYSIZE |
983 OP_XVALID_LAZYBLOCKS);
985 mdc_close_pack(&req->rq_pill, op_data);
987 req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
988 obd->u.cli.cl_default_mds_easize);
990 ptlrpc_request_set_replen(req);
992 ptlrpc_get_mod_rpc_slot(req);
993 rc = ptlrpc_queue_wait(req);
994 ptlrpc_put_mod_rpc_slot(req);
996 if (req->rq_repmsg == NULL) {
997 CDEBUG(D_RPCTRACE, "request %p failed to send: rc = %d\n", req,
1000 rc = req->rq_status ?: -EIO;
1001 } else if (rc == 0 || rc == -EAGAIN) {
1002 struct mdt_body *body;
1004 rc = lustre_msg_get_status(req->rq_repmsg);
1005 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
1006 DEBUG_REQ(D_ERROR, req,
1007 "type = PTL_RPC_MSG_ERR: rc = %d", rc);
1011 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
1014 } else if (rc == -ESTALE) {
1016 * it can be allowed error after 3633 if open was committed and
1017 * server failed before close was sent. Let's check if mod
1018 * exists and return no error in that case
1021 DEBUG_REQ(D_HA, req, "Reset ESTALE = %d", rc);
1022 LASSERT(mod->mod_open_req != NULL);
1023 if (mod->mod_open_req->rq_committed)
1031 mod->mod_close_req = NULL;
1032 /* Since now, mod is accessed through open_req only,
1033 * thus close req does not keep a reference on mod anymore. */
1038 RETURN(rc < 0 ? rc : saved_rc);
1041 static int mdc_getpage(struct obd_export *exp, const struct lu_fid *fid,
1042 u64 offset, struct page **pages, int npages,
1043 struct ptlrpc_request **request)
1045 struct ptlrpc_request *req;
1046 struct ptlrpc_bulk_desc *desc;
1055 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_READPAGE);
1059 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_READPAGE);
1061 ptlrpc_request_free(req);
1065 req->rq_request_portal = MDS_READPAGE_PORTAL;
1066 ptlrpc_at_set_req_timeout(req);
1068 desc = ptlrpc_prep_bulk_imp(req, npages, 1,
1069 PTLRPC_BULK_PUT_SINK,
1071 &ptlrpc_bulk_kiov_pin_ops);
1073 ptlrpc_req_finished(req);
1077 /* NB req now owns desc and will free it when it gets freed */
1078 for (i = 0; i < npages; i++)
1079 desc->bd_frag_ops->add_kiov_frag(desc, pages[i], 0,
1082 mdc_readdir_pack(&req->rq_pill, offset, PAGE_SIZE * npages, fid);
1084 ptlrpc_request_set_replen(req);
1085 rc = ptlrpc_queue_wait(req);
1087 ptlrpc_req_finished(req);
1088 if (rc != -ETIMEDOUT)
1092 if (!client_should_resend(resends, &exp->exp_obd->u.cli)) {
1093 CERROR("%s: too many resend retries: rc = %d\n",
1094 exp->exp_obd->obd_name, -EIO);
1098 /* If a signal interrupts then the timeout returned will
1099 * not be zero. In that case return -EINTR
1101 if (msleep_interruptible(resends * 1000))
1107 rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk,
1108 req->rq_bulk->bd_nob_transferred);
1110 ptlrpc_req_finished(req);
1114 if (req->rq_bulk->bd_nob_transferred & ~LU_PAGE_MASK) {
1115 CERROR("%s: unexpected bytes transferred: %d (%ld expected)\n",
1116 exp->exp_obd->obd_name, req->rq_bulk->bd_nob_transferred,
1117 PAGE_SIZE * npages);
1118 ptlrpc_req_finished(req);
1126 static void mdc_release_page(struct page *page, int remove)
1130 if (likely(page->mapping != NULL))
1131 delete_from_page_cache(page);
1137 static struct page *mdc_page_locate(struct address_space *mapping, __u64 *hash,
1138 __u64 *start, __u64 *end, int hash64)
1141 * Complement of hash is used as an index so that
1142 * radix_tree_gang_lookup() can be used to find a page with starting
1143 * hash _smaller_ than one we are looking for.
1145 unsigned long offset = hash_x_index(*hash, hash64);
1147 unsigned long flags;
1150 ll_xa_lock_irqsave(&mapping->i_pages, flags);
1151 found = radix_tree_gang_lookup(&mapping->page_tree,
1152 (void **)&page, offset, 1);
1153 if (found > 0 && !ll_xa_is_value(page)) {
1154 struct lu_dirpage *dp;
1157 ll_xa_unlock_irqrestore(&mapping->i_pages, flags);
1159 * In contrast to find_lock_page() we are sure that directory
1160 * page cannot be truncated (while DLM lock is held) and,
1161 * hence, can avoid restart.
1163 * In fact, page cannot be locked here at all, because
1164 * mdc_read_page_remote does synchronous io.
1166 wait_on_page_locked(page);
1167 if (PageUptodate(page)) {
1169 if (BITS_PER_LONG == 32 && hash64) {
1170 *start = le64_to_cpu(dp->ldp_hash_start) >> 32;
1171 *end = le64_to_cpu(dp->ldp_hash_end) >> 32;
1172 *hash = *hash >> 32;
1174 *start = le64_to_cpu(dp->ldp_hash_start);
1175 *end = le64_to_cpu(dp->ldp_hash_end);
1177 if (unlikely(*start == 1 && *hash == 0))
1180 LASSERTF(*start <= *hash, "start = %#llx"
1181 ",end = %#llx,hash = %#llx\n",
1182 *start, *end, *hash);
1183 CDEBUG(D_VFSTRACE, "offset %lx [%#llx %#llx],"
1184 " hash %#llx\n", offset, *start, *end, *hash);
1187 mdc_release_page(page, 0);
1189 } else if (*end != *start && *hash == *end) {
1191 * upon hash collision, remove this page,
1192 * otherwise put page reference, and
1193 * mdc_read_page_remote() will issue RPC to
1194 * fetch the page we want.
1197 mdc_release_page(page,
1198 le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
1203 page = ERR_PTR(-EIO);
1206 ll_xa_unlock_irqrestore(&mapping->i_pages, flags);
1213 * Adjust a set of pages, each page containing an array of lu_dirpages,
1214 * so that each page can be used as a single logical lu_dirpage.
1216 * A lu_dirpage is laid out as follows, where s = ldp_hash_start,
1217 * e = ldp_hash_end, f = ldp_flags, p = padding, and each "ent" is a
1218 * struct lu_dirent. It has size up to LU_PAGE_SIZE. The ldp_hash_end
1219 * value is used as a cookie to request the next lu_dirpage in a
1220 * directory listing that spans multiple pages (two in this example):
1223 * .|--------v------- -----.
1224 * |s|e|f|p|ent|ent| ... |ent|
1225 * '--|-------------- -----' Each PAGE contains a single
1226 * '------. lu_dirpage.
1227 * .---------v------- -----.
1228 * |s|e|f|p|ent| 0 | ... | 0 |
1229 * '----------------- -----'
1231 * However, on hosts where the native VM page size (PAGE_SIZE) is
1232 * larger than LU_PAGE_SIZE, a single host page may contain multiple
1233 * lu_dirpages. After reading the lu_dirpages from the MDS, the
1234 * ldp_hash_end of the first lu_dirpage refers to the one immediately
1235 * after it in the same PAGE (arrows simplified for brevity, but
1236 * in general e0==s1, e1==s2, etc.):
1238 * .-------------------- -----.
1239 * |s0|e0|f0|p|ent|ent| ... |ent|
1240 * |---v---------------- -----|
1241 * |s1|e1|f1|p|ent|ent| ... |ent|
1242 * |---v---------------- -----| Here, each PAGE contains
1243 * ... multiple lu_dirpages.
1244 * |---v---------------- -----|
1245 * |s'|e'|f'|p|ent|ent| ... |ent|
1246 * '---|---------------- -----'
1248 * .----------------------------.
1251 * This structure is transformed into a single logical lu_dirpage as follows:
1253 * - Replace e0 with e' so the request for the next lu_dirpage gets the page
1254 * labeled 'next PAGE'.
1256 * - Copy the LDF_COLLIDE flag from f' to f0 to correctly reflect whether
1257 * a hash collision with the next page exists.
1259 * - Adjust the lde_reclen of the ending entry of each lu_dirpage to span
1260 * to the first entry of the next lu_dirpage.
1262 #if PAGE_SIZE > LU_PAGE_SIZE
1263 static void mdc_adjust_dirpages(struct page **pages, int cfs_pgs, int lu_pgs)
1267 for (i = 0; i < cfs_pgs; i++) {
1268 struct lu_dirpage *dp = kmap(pages[i]);
1269 struct lu_dirpage *first = dp;
1270 struct lu_dirent *end_dirent = NULL;
1271 struct lu_dirent *ent;
1272 __u64 hash_end = dp->ldp_hash_end;
1273 __u32 flags = dp->ldp_flags;
1275 while (--lu_pgs > 0) {
1276 ent = lu_dirent_start(dp);
1277 for (end_dirent = ent; ent != NULL;
1278 end_dirent = ent, ent = lu_dirent_next(ent));
1280 /* Advance dp to next lu_dirpage. */
1281 dp = (struct lu_dirpage *)((char *)dp + LU_PAGE_SIZE);
1283 /* Check if we've reached the end of the PAGE. */
1284 if (!((unsigned long)dp & ~PAGE_MASK))
1287 /* Save the hash and flags of this lu_dirpage. */
1288 hash_end = dp->ldp_hash_end;
1289 flags = dp->ldp_flags;
1291 /* Check if lu_dirpage contains no entries. */
1292 if (end_dirent == NULL)
1295 /* Enlarge the end entry lde_reclen from 0 to
1296 * first entry of next lu_dirpage. */
1297 LASSERT(le16_to_cpu(end_dirent->lde_reclen) == 0);
1298 end_dirent->lde_reclen =
1299 cpu_to_le16((char *)(dp->ldp_entries) -
1300 (char *)end_dirent);
1303 first->ldp_hash_end = hash_end;
1304 first->ldp_flags &= ~cpu_to_le32(LDF_COLLIDE);
1305 first->ldp_flags |= flags & cpu_to_le32(LDF_COLLIDE);
1309 LASSERTF(lu_pgs == 0, "left = %d\n", lu_pgs);
1312 #define mdc_adjust_dirpages(pages, cfs_pgs, lu_pgs) do {} while (0)
1313 #endif /* PAGE_SIZE > LU_PAGE_SIZE */
1315 /* parameters for readdir page */
1316 struct readpage_param {
1317 struct md_op_data *rp_mod;
1320 struct obd_export *rp_exp;
1324 * Read pages from server.
1326 * Page in MDS_READPAGE RPC is packed in LU_PAGE_SIZE, and each page contains
1327 * a header lu_dirpage which describes the start/end hash, and whether this
1328 * page is empty (contains no dir entry) or hash collide with next page.
1329 * After client receives reply, several pages will be integrated into dir page
1330 * in PAGE_SIZE (if PAGE_SIZE greater than LU_PAGE_SIZE), and the
1331 * lu_dirpage for this integrated page will be adjusted.
1333 static int mdc_read_page_remote(void *data, struct page *page0)
1335 struct readpage_param *rp = data;
1336 struct page **page_pool;
1338 struct lu_dirpage *dp;
1339 struct md_op_data *op_data = rp->rp_mod;
1340 struct ptlrpc_request *req;
1342 struct inode *inode;
1344 int rd_pgs = 0; /* number of pages actually read */
1350 max_pages = rp->rp_exp->exp_obd->u.cli.cl_max_pages_per_rpc;
1351 inode = op_data->op_data;
1352 fid = &op_data->op_fid1;
1353 LASSERT(inode != NULL);
1355 OBD_ALLOC_PTR_ARRAY_LARGE(page_pool, max_pages);
1356 if (page_pool != NULL) {
1357 page_pool[0] = page0;
1363 for (npages = 1; npages < max_pages; npages++) {
1364 page = page_cache_alloc(inode->i_mapping);
1367 page_pool[npages] = page;
1370 rc = mdc_getpage(rp->rp_exp, fid, rp->rp_off, page_pool, npages, &req);
1372 /* page0 is special, which was added into page cache early */
1373 delete_from_page_cache(page0);
1377 rd_pgs = (req->rq_bulk->bd_nob_transferred + PAGE_SIZE - 1) >>
1379 lu_pgs = req->rq_bulk->bd_nob_transferred >> LU_PAGE_SHIFT;
1380 LASSERT(!(req->rq_bulk->bd_nob_transferred & ~LU_PAGE_MASK));
1382 CDEBUG(D_INODE, "read %d(%d) pages\n", rd_pgs, lu_pgs);
1384 mdc_adjust_dirpages(page_pool, rd_pgs, lu_pgs);
1386 SetPageUptodate(page0);
1390 ptlrpc_req_finished(req);
1391 CDEBUG(D_CACHE, "read %d/%d pages\n", rd_pgs, npages);
1392 for (i = 1; i < npages; i++) {
1393 unsigned long offset;
1397 page = page_pool[i];
1399 if (rc < 0 || i >= rd_pgs) {
1404 SetPageUptodate(page);
1407 hash = le64_to_cpu(dp->ldp_hash_start);
1410 offset = hash_x_index(hash, rp->rp_hash64);
1412 prefetchw(&page->flags);
1413 ret = add_to_page_cache_lru(page, inode->i_mapping, offset,
1418 CDEBUG(D_VFSTRACE, "page %lu add to page cache failed:"
1419 " rc = %d\n", offset, ret);
1423 if (page_pool != &page0)
1424 OBD_FREE_PTR_ARRAY_LARGE(page_pool, max_pages);
1430 * Read dir page from cache first, if it can not find it, read it from
1431 * server and add into the cache.
1433 * \param[in] exp MDC export
1434 * \param[in] op_data client MD stack parameters, transfering parameters
1435 * between different layers on client MD stack.
1436 * \param[in] mrinfo callback required for ldlm lock enqueue during
1438 * \param[in] hash_offset the hash offset of the page to be read
1439 * \param[in] ppage the page to be read
1441 * retval = 0 get the page successfully
1442 * errno(<0) get the page failed
1444 static int mdc_read_page(struct obd_export *exp, struct md_op_data *op_data,
1445 struct md_readdir_info *mrinfo, __u64 hash_offset,
1446 struct page **ppage)
1448 struct lookup_intent it = { .it_op = IT_READDIR };
1450 struct inode *dir = op_data->op_data;
1451 struct address_space *mapping;
1452 struct lu_dirpage *dp;
1455 struct lustre_handle lockh;
1456 struct ptlrpc_request *enq_req = NULL;
1457 struct readpage_param rp_param;
1464 LASSERT(dir != NULL);
1465 mapping = dir->i_mapping;
1467 rc = mdc_intent_lock(exp, op_data, &it, &enq_req,
1468 mrinfo->mr_blocking_ast, 0);
1469 if (enq_req != NULL)
1470 ptlrpc_req_finished(enq_req);
1473 CERROR("%s: "DFID" lock enqueue fails: rc = %d\n",
1474 exp->exp_obd->obd_name, PFID(&op_data->op_fid1), rc);
1479 lockh.cookie = it.it_lock_handle;
1480 mdc_set_lock_data(exp, &lockh, dir, NULL);
1482 rp_param.rp_off = hash_offset;
1483 rp_param.rp_hash64 = op_data->op_cli_flags & CLI_HASH64;
1484 page = mdc_page_locate(mapping, &rp_param.rp_off, &start, &end,
1485 rp_param.rp_hash64);
1487 CERROR("%s: dir page locate: "DFID" at %llu: rc %ld\n",
1488 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1489 rp_param.rp_off, PTR_ERR(page));
1490 GOTO(out_unlock, rc = PTR_ERR(page));
1491 } else if (page != NULL) {
1493 * XXX nikita: not entirely correct handling of a corner case:
1494 * suppose hash chain of entries with hash value HASH crosses
1495 * border between pages P0 and P1. First both P0 and P1 are
1496 * cached, seekdir() is called for some entry from the P0 part
1497 * of the chain. Later P0 goes out of cache. telldir(HASH)
1498 * happens and finds P1, as it starts with matching hash
1499 * value. Remaining entries from P0 part of the chain are
1500 * skipped. (Is that really a bug?)
1502 * Possible solutions: 0. don't cache P1 is such case, handle
1503 * it as an "overflow" page. 1. invalidate all pages at
1504 * once. 2. use HASH|1 as an index for P1.
1506 GOTO(hash_collision, page);
1509 rp_param.rp_exp = exp;
1510 rp_param.rp_mod = op_data;
1511 page = read_cache_page(mapping,
1512 hash_x_index(rp_param.rp_off,
1513 rp_param.rp_hash64),
1514 mdc_read_page_remote, &rp_param);
1516 CDEBUG(D_INFO, "%s: read cache page: "DFID" at %llu: %ld\n",
1517 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1518 rp_param.rp_off, PTR_ERR(page));
1519 GOTO(out_unlock, rc = PTR_ERR(page));
1522 wait_on_page_locked(page);
1524 if (!PageUptodate(page)) {
1525 CERROR("%s: page not updated: "DFID" at %llu: rc %d\n",
1526 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1527 rp_param.rp_off, -5);
1530 if (!PageChecked(page))
1531 SetPageChecked(page);
1532 if (PageError(page)) {
1533 CERROR("%s: page error: "DFID" at %llu: rc %d\n",
1534 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1535 rp_param.rp_off, -5);
1540 dp = page_address(page);
1541 if (BITS_PER_LONG == 32 && rp_param.rp_hash64) {
1542 start = le64_to_cpu(dp->ldp_hash_start) >> 32;
1543 end = le64_to_cpu(dp->ldp_hash_end) >> 32;
1544 rp_param.rp_off = hash_offset >> 32;
1546 start = le64_to_cpu(dp->ldp_hash_start);
1547 end = le64_to_cpu(dp->ldp_hash_end);
1548 rp_param.rp_off = hash_offset;
1551 LASSERT(start == rp_param.rp_off);
1552 CWARN("Page-wide hash collision: %#lx\n", (unsigned long)end);
1553 #if BITS_PER_LONG == 32
1554 CWARN("Real page-wide hash collision at [%llu %llu] with "
1555 "hash %llu\n", le64_to_cpu(dp->ldp_hash_start),
1556 le64_to_cpu(dp->ldp_hash_end), hash_offset);
1560 * Fetch whole overflow chain...
1568 ldlm_lock_decref(&lockh, it.it_lock_mode);
1572 mdc_release_page(page, 1);
1577 static int mdc_statfs_interpret(const struct lu_env *env,
1578 struct ptlrpc_request *req, void *args, int rc)
1580 struct obd_info *oinfo = args;
1581 struct obd_statfs *osfs;
1584 osfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
1588 oinfo->oi_osfs = osfs;
1590 CDEBUG(D_CACHE, "blocks=%llu free=%llu avail=%llu "
1591 "objects=%llu free=%llu state=%x\n",
1592 osfs->os_blocks, osfs->os_bfree, osfs->os_bavail,
1593 osfs->os_files, osfs->os_ffree, osfs->os_state);
1596 oinfo->oi_cb_up(oinfo, rc);
1601 static int mdc_statfs_async(struct obd_export *exp,
1602 struct obd_info *oinfo, time64_t max_age,
1603 struct ptlrpc_request_set *unused)
1605 struct ptlrpc_request *req;
1606 struct obd_info *aa;
1608 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp), &RQF_MDS_STATFS,
1609 LUSTRE_MDS_VERSION, MDS_STATFS);
1613 ptlrpc_request_set_replen(req);
1614 req->rq_interpret_reply = mdc_statfs_interpret;
1616 aa = ptlrpc_req_async_args(aa, req);
1619 ptlrpcd_add_req(req);
1624 static int mdc_statfs(const struct lu_env *env,
1625 struct obd_export *exp, struct obd_statfs *osfs,
1626 time64_t max_age, __u32 flags)
1628 struct obd_device *obd = class_exp2obd(exp);
1629 struct req_format *fmt;
1630 struct ptlrpc_request *req;
1631 struct obd_statfs *msfs;
1632 struct obd_import *imp, *imp0;
1637 * Since the request might also come from lprocfs, so we need
1638 * sync this with client_disconnect_export Bug15684
1640 with_imp_locked(obd, imp0, rc)
1641 imp = class_import_get(imp0);
1645 fmt = &RQF_MDS_STATFS;
1646 if ((exp_connect_flags2(exp) & OBD_CONNECT2_SUM_STATFS) &&
1647 (flags & OBD_STATFS_SUM))
1648 fmt = &RQF_MDS_STATFS_NEW;
1649 req = ptlrpc_request_alloc_pack(imp, fmt, LUSTRE_MDS_VERSION,
1652 GOTO(output, rc = -ENOMEM);
1653 req->rq_allow_intr = 1;
1655 if ((flags & OBD_STATFS_SUM) &&
1656 (exp_connect_flags2(exp) & OBD_CONNECT2_SUM_STATFS)) {
1657 /* request aggregated states */
1658 struct mdt_body *body;
1660 body = req_capsule_client_get(&req->rq_pill, &RMF_MDT_BODY);
1662 GOTO(out, rc = -EPROTO);
1663 body->mbo_valid = OBD_MD_FLAGSTATFS;
1666 ptlrpc_request_set_replen(req);
1668 if (flags & OBD_STATFS_NODELAY) {
1669 /* procfs requests not want stay in wait for avoid deadlock */
1670 req->rq_no_resend = 1;
1671 req->rq_no_delay = 1;
1674 rc = ptlrpc_queue_wait(req);
1676 /* check connection error first */
1677 if (imp->imp_connect_error)
1678 rc = imp->imp_connect_error;
1682 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
1684 GOTO(out, rc = -EPROTO);
1689 ptlrpc_req_finished(req);
1691 class_import_put(imp);
1695 static int mdc_ioc_fid2path(struct obd_export *exp, struct getinfo_fid2path *gf)
1697 __u32 keylen, vallen;
1701 if (gf->gf_pathlen > PATH_MAX)
1702 RETURN(-ENAMETOOLONG);
1703 if (gf->gf_pathlen < 2)
1706 /* Key is KEY_FID2PATH + getinfo_fid2path description */
1707 keylen = cfs_size_round(sizeof(KEY_FID2PATH) + sizeof(*gf) +
1708 sizeof(struct lu_fid));
1709 OBD_ALLOC(key, keylen);
1712 memcpy(key, KEY_FID2PATH, sizeof(KEY_FID2PATH));
1713 memcpy(key + cfs_size_round(sizeof(KEY_FID2PATH)), gf, sizeof(*gf));
1714 memcpy(key + cfs_size_round(sizeof(KEY_FID2PATH)) + sizeof(*gf),
1715 gf->gf_u.gf_root_fid, sizeof(struct lu_fid));
1716 CDEBUG(D_IOCTL, "path get "DFID" from %llu #%d\n",
1717 PFID(&gf->gf_fid), gf->gf_recno, gf->gf_linkno);
1719 if (!fid_is_sane(&gf->gf_fid))
1720 GOTO(out, rc = -EINVAL);
1722 /* Val is struct getinfo_fid2path result plus path */
1723 vallen = sizeof(*gf) + gf->gf_pathlen;
1725 rc = obd_get_info(NULL, exp, keylen, key, &vallen, gf);
1726 if (rc != 0 && rc != -EREMOTE)
1729 if (vallen <= sizeof(*gf))
1730 GOTO(out, rc = -EPROTO);
1731 if (vallen > sizeof(*gf) + gf->gf_pathlen)
1732 GOTO(out, rc = -EOVERFLOW);
1734 CDEBUG(D_IOCTL, "path got "DFID" from %llu #%d: %s\n",
1735 PFID(&gf->gf_fid), gf->gf_recno, gf->gf_linkno,
1736 gf->gf_pathlen < 512 ? gf->gf_u.gf_path :
1737 /* only log the last 512 characters of the path */
1738 gf->gf_u.gf_path + gf->gf_pathlen - 512);
1741 OBD_FREE(key, keylen);
1745 static int mdc_ioc_hsm_progress(struct obd_export *exp,
1746 struct hsm_progress_kernel *hpk)
1748 struct obd_import *imp = class_exp2cliimp(exp);
1749 struct hsm_progress_kernel *req_hpk;
1750 struct ptlrpc_request *req;
1754 req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_HSM_PROGRESS,
1755 LUSTRE_MDS_VERSION, MDS_HSM_PROGRESS);
1757 GOTO(out, rc = -ENOMEM);
1759 mdc_pack_body(&req->rq_pill, NULL, 0, 0, -1, 0);
1761 /* Copy hsm_progress struct */
1762 req_hpk = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_PROGRESS);
1763 if (req_hpk == NULL)
1764 GOTO(out, rc = -EPROTO);
1767 req_hpk->hpk_errval = lustre_errno_hton(hpk->hpk_errval);
1769 ptlrpc_request_set_replen(req);
1771 ptlrpc_get_mod_rpc_slot(req);
1772 rc = ptlrpc_queue_wait(req);
1773 ptlrpc_put_mod_rpc_slot(req);
1777 ptlrpc_req_finished(req);
1781 * Send hsm_ct_register to MDS
1783 * \param[in] imp import
1784 * \param[in] archive_count if in bitmap format, it is the bitmap,
1785 * else it is the count of archive_ids
1786 * \param[in] archives if in bitmap format, it is NULL,
1787 * else it is archive_id lists
1789 static int mdc_ioc_hsm_ct_register(struct obd_import *imp, __u32 archive_count,
1792 struct ptlrpc_request *req;
1793 __u32 *archive_array;
1794 size_t archives_size;
1798 req = ptlrpc_request_alloc(imp, &RQF_MDS_HSM_CT_REGISTER);
1802 if (archives != NULL)
1803 archives_size = sizeof(*archive_array) * archive_count;
1805 archives_size = sizeof(archive_count);
1807 req_capsule_set_size(&req->rq_pill, &RMF_MDS_HSM_ARCHIVE,
1808 RCL_CLIENT, archives_size);
1810 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_CT_REGISTER);
1812 ptlrpc_request_free(req);
1816 mdc_pack_body(&req->rq_pill, NULL, 0, 0, -1, 0);
1818 archive_array = req_capsule_client_get(&req->rq_pill,
1819 &RMF_MDS_HSM_ARCHIVE);
1820 if (archive_array == NULL)
1821 GOTO(out, rc = -EPROTO);
1823 if (archives != NULL)
1824 memcpy(archive_array, archives, archives_size);
1826 *archive_array = archive_count;
1828 ptlrpc_request_set_replen(req);
1829 req->rq_no_resend = 1;
1831 rc = mdc_queue_wait(req);
1834 ptlrpc_req_finished(req);
1838 static int mdc_ioc_hsm_current_action(struct obd_export *exp,
1839 struct md_op_data *op_data)
1841 struct hsm_current_action *hca = op_data->op_data;
1842 struct hsm_current_action *req_hca;
1843 struct ptlrpc_request *req;
1847 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1848 &RQF_MDS_HSM_ACTION);
1852 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_ACTION);
1854 ptlrpc_request_free(req);
1858 mdc_pack_body(&req->rq_pill, &op_data->op_fid1, 0, 0,
1859 op_data->op_suppgids[0], 0);
1861 ptlrpc_request_set_replen(req);
1863 rc = mdc_queue_wait(req);
1867 req_hca = req_capsule_server_get(&req->rq_pill,
1868 &RMF_MDS_HSM_CURRENT_ACTION);
1869 if (req_hca == NULL)
1870 GOTO(out, rc = -EPROTO);
1876 ptlrpc_req_finished(req);
1880 static int mdc_ioc_hsm_ct_unregister(struct obd_import *imp)
1882 struct ptlrpc_request *req;
1886 req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_HSM_CT_UNREGISTER,
1888 MDS_HSM_CT_UNREGISTER);
1890 GOTO(out, rc = -ENOMEM);
1892 mdc_pack_body(&req->rq_pill, NULL, 0, 0, -1, 0);
1894 ptlrpc_request_set_replen(req);
1896 rc = mdc_queue_wait(req);
1899 ptlrpc_req_finished(req);
1903 static int mdc_ioc_hsm_state_get(struct obd_export *exp,
1904 struct md_op_data *op_data)
1906 struct hsm_user_state *hus = op_data->op_data;
1907 struct hsm_user_state *req_hus;
1908 struct ptlrpc_request *req;
1912 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1913 &RQF_MDS_HSM_STATE_GET);
1917 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_STATE_GET);
1919 ptlrpc_request_free(req);
1923 mdc_pack_body(&req->rq_pill, &op_data->op_fid1, 0, 0,
1924 op_data->op_suppgids[0], 0);
1926 ptlrpc_request_set_replen(req);
1928 rc = mdc_queue_wait(req);
1932 req_hus = req_capsule_server_get(&req->rq_pill, &RMF_HSM_USER_STATE);
1933 if (req_hus == NULL)
1934 GOTO(out, rc = -EPROTO);
1940 ptlrpc_req_finished(req);
1944 static int mdc_ioc_hsm_state_set(struct obd_export *exp,
1945 struct md_op_data *op_data)
1947 struct hsm_state_set *hss = op_data->op_data;
1948 struct hsm_state_set *req_hss;
1949 struct ptlrpc_request *req;
1953 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1954 &RQF_MDS_HSM_STATE_SET);
1958 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_STATE_SET);
1960 ptlrpc_request_free(req);
1964 mdc_pack_body(&req->rq_pill, &op_data->op_fid1, 0, 0,
1965 op_data->op_suppgids[0], 0);
1968 req_hss = req_capsule_client_get(&req->rq_pill, &RMF_HSM_STATE_SET);
1969 if (req_hss == NULL)
1970 GOTO(out, rc = -EPROTO);
1973 ptlrpc_request_set_replen(req);
1975 ptlrpc_get_mod_rpc_slot(req);
1976 rc = ptlrpc_queue_wait(req);
1977 ptlrpc_put_mod_rpc_slot(req);
1981 ptlrpc_req_finished(req);
1985 /* For RESTORE and RELEASE the mdt will take EX lock on the file layout.
1986 * So we can use early cancel on client side locks for that resource.
1988 static inline int mdc_hsm_request_lock_to_cancel(struct obd_export *exp,
1989 struct hsm_user_request *hur,
1990 struct list_head *cancels)
1992 struct hsm_user_item *hui = &hur->hur_user_item[0];
1993 struct hsm_request *req_hr = &hur->hur_request;
1997 if (req_hr->hr_action != HUA_RESTORE &&
1998 req_hr->hr_action != HUA_RELEASE)
2001 for (i = 0; i < req_hr->hr_itemcount; i++, hui++) {
2002 if (!fid_is_sane(&hui->hui_fid))
2004 count += mdc_resource_get_unused(exp, &hui->hui_fid, cancels,
2005 LCK_EX, MDS_INODELOCK_LAYOUT);
2011 static int mdc_ioc_hsm_request(struct obd_export *exp,
2012 struct hsm_user_request *hur)
2014 struct obd_import *imp = class_exp2cliimp(exp);
2015 struct ptlrpc_request *req;
2016 struct hsm_request *req_hr;
2017 struct hsm_user_item *req_hui;
2024 req = ptlrpc_request_alloc(imp, &RQF_MDS_HSM_REQUEST);
2028 req_capsule_set_size(&req->rq_pill, &RMF_MDS_HSM_USER_ITEM, RCL_CLIENT,
2029 hur->hur_request.hr_itemcount
2030 * sizeof(struct hsm_user_item));
2031 req_capsule_set_size(&req->rq_pill, &RMF_GENERIC_DATA, RCL_CLIENT,
2032 hur->hur_request.hr_data_len);
2034 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_REQUEST);
2036 ptlrpc_request_free(req);
2040 /* Cancel existing locks */
2041 count = mdc_hsm_request_lock_to_cancel(exp, hur, &cancels);
2042 ldlm_cli_cancel_list(&cancels, count, NULL, 0);
2043 mdc_pack_body(&req->rq_pill, NULL, 0, 0, -1, 0);
2045 /* Copy hsm_request struct */
2046 req_hr = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_REQUEST);
2048 GOTO(out, rc = -EPROTO);
2049 *req_hr = hur->hur_request;
2051 /* Copy hsm_user_item structs */
2052 req_hui = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_USER_ITEM);
2053 if (req_hui == NULL)
2054 GOTO(out, rc = -EPROTO);
2055 memcpy(req_hui, hur->hur_user_item,
2056 hur->hur_request.hr_itemcount * sizeof(struct hsm_user_item));
2058 /* Copy opaque field */
2059 req_opaque = req_capsule_client_get(&req->rq_pill, &RMF_GENERIC_DATA);
2060 if (req_opaque == NULL)
2061 GOTO(out, rc = -EPROTO);
2062 memcpy(req_opaque, hur_data(hur), hur->hur_request.hr_data_len);
2064 ptlrpc_request_set_replen(req);
2066 ptlrpc_get_mod_rpc_slot(req);
2067 rc = ptlrpc_queue_wait(req);
2068 ptlrpc_put_mod_rpc_slot(req);
2073 ptlrpc_req_finished(req);
2077 static int mdc_ioc_hsm_ct_start(struct obd_export *exp,
2078 struct lustre_kernelcomm *lk);
2080 static int mdc_quotactl(struct obd_device *unused, struct obd_export *exp,
2081 struct obd_quotactl *oqctl)
2083 struct ptlrpc_request *req;
2084 struct obd_quotactl *oqc;
2088 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_QUOTACTL);
2093 if (LUSTRE_Q_CMD_IS_POOL(oqctl->qc_cmd))
2094 req_capsule_set_size(&req->rq_pill,
2097 sizeof(*oqc) + LOV_MAXPOOLNAME + 1);
2099 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION,
2102 ptlrpc_request_free(req);
2106 oqc = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
2107 QCTL_COPY(oqc, oqctl);
2109 ptlrpc_request_set_replen(req);
2110 ptlrpc_at_set_req_timeout(req);
2112 rc = ptlrpc_queue_wait(req);
2114 CERROR("%s: ptlrpc_queue_wait failed: rc = %d\n",
2115 exp->exp_obd->obd_name, rc);
2119 if (req->rq_repmsg &&
2120 (oqc = req_capsule_server_get(&req->rq_pill, &RMF_OBD_QUOTACTL))) {
2121 QCTL_COPY(oqctl, oqc);
2124 CERROR("%s: cannot unpack obd_quotactl: rc = %d\n",
2125 exp->exp_obd->obd_name, rc);
2128 ptlrpc_req_finished(req);
2133 static int mdc_ioc_swap_layouts(struct obd_export *exp,
2134 struct md_op_data *op_data)
2137 struct ptlrpc_request *req;
2139 struct mdc_swap_layouts *msl, *payload;
2142 msl = op_data->op_data;
2144 /* When the MDT will get the MDS_SWAP_LAYOUTS RPC the
2145 * first thing it will do is to cancel the 2 layout
2146 * locks held by this client.
2147 * So the client must cancel its layout locks on the 2 fids
2148 * with the request RPC to avoid extra RPC round trips.
2150 count = mdc_resource_get_unused(exp, &op_data->op_fid1, &cancels,
2151 LCK_EX, MDS_INODELOCK_LAYOUT |
2152 MDS_INODELOCK_XATTR);
2153 count += mdc_resource_get_unused(exp, &op_data->op_fid2, &cancels,
2154 LCK_EX, MDS_INODELOCK_LAYOUT |
2155 MDS_INODELOCK_XATTR);
2157 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
2158 &RQF_MDS_SWAP_LAYOUTS);
2160 ldlm_lock_list_put(&cancels, l_bl_ast, count);
2164 rc = mdc_prep_elc_req(exp, req, MDS_SWAP_LAYOUTS, &cancels, count);
2166 ptlrpc_request_free(req);
2170 mdc_swap_layouts_pack(&req->rq_pill, op_data);
2172 payload = req_capsule_client_get(&req->rq_pill, &RMF_SWAP_LAYOUTS);
2177 ptlrpc_request_set_replen(req);
2179 rc = ptlrpc_queue_wait(req);
2185 ptlrpc_req_finished(req);
2189 static int mdc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2190 void *karg, void __user *uarg)
2192 struct obd_device *obd = exp->exp_obd;
2193 struct obd_ioctl_data *data = karg;
2194 struct obd_import *imp = obd->u.cli.cl_import;
2198 if (!try_module_get(THIS_MODULE)) {
2199 CERROR("%s: cannot get module '%s'\n", obd->obd_name,
2200 module_name(THIS_MODULE));
2204 case OBD_IOC_FID2PATH:
2205 rc = mdc_ioc_fid2path(exp, karg);
2207 case LL_IOC_HSM_CT_START:
2208 rc = mdc_ioc_hsm_ct_start(exp, karg);
2209 /* ignore if it was already registered on this MDS. */
2213 case LL_IOC_HSM_PROGRESS:
2214 rc = mdc_ioc_hsm_progress(exp, karg);
2216 case LL_IOC_HSM_STATE_GET:
2217 rc = mdc_ioc_hsm_state_get(exp, karg);
2219 case LL_IOC_HSM_STATE_SET:
2220 rc = mdc_ioc_hsm_state_set(exp, karg);
2222 case LL_IOC_HSM_ACTION:
2223 rc = mdc_ioc_hsm_current_action(exp, karg);
2225 case LL_IOC_HSM_REQUEST:
2226 rc = mdc_ioc_hsm_request(exp, karg);
2228 case OBD_IOC_CLIENT_RECOVER:
2229 rc = ptlrpc_recover_import(imp, data->ioc_inlbuf1, 0);
2233 case IOC_OSC_SET_ACTIVE:
2234 rc = ptlrpc_set_import_active(imp, data->ioc_offset);
2237 * Normally IOC_OBD_STATFS, OBD_IOC_QUOTACTL iocontrol are handled by
2238 * LMV instead of MDC. But when the cluster is upgraded from 1.8,
2239 * there'd be no LMV layer thus we might be called here. Eventually
2240 * this code should be removed.
2243 case IOC_OBD_STATFS: {
2244 struct obd_statfs stat_buf = {0};
2246 if (*((__u32 *) data->ioc_inlbuf2) != 0)
2247 GOTO(out, rc = -ENODEV);
2250 if (copy_to_user(data->ioc_pbuf2, obd2cli_tgt(obd),
2251 min((int)data->ioc_plen2,
2252 (int)sizeof(struct obd_uuid))))
2253 GOTO(out, rc = -EFAULT);
2255 rc = mdc_statfs(NULL, obd->obd_self_export, &stat_buf,
2256 ktime_get_seconds() - OBD_STATFS_CACHE_SECONDS,
2261 if (copy_to_user(data->ioc_pbuf1, &stat_buf,
2262 min((int) data->ioc_plen1,
2263 (int) sizeof(stat_buf))))
2264 GOTO(out, rc = -EFAULT);
2268 case OBD_IOC_QUOTACTL: {
2269 struct if_quotactl *qctl = karg;
2270 struct obd_quotactl *oqctl;
2272 OBD_ALLOC_PTR(oqctl);
2274 GOTO(out, rc = -ENOMEM);
2276 QCTL_COPY(oqctl, qctl);
2277 rc = obd_quotactl(exp, oqctl);
2279 QCTL_COPY(qctl, oqctl);
2280 qctl->qc_valid = QC_MDTIDX;
2281 qctl->obd_uuid = obd->u.cli.cl_target_uuid;
2284 OBD_FREE_PTR(oqctl);
2287 case LL_IOC_GET_CONNECT_FLAGS:
2288 if (copy_to_user(uarg, exp_connect_flags_ptr(exp),
2289 sizeof(*exp_connect_flags_ptr(exp))))
2290 GOTO(out, rc = -EFAULT);
2293 case LL_IOC_LOV_SWAP_LAYOUTS:
2294 rc = mdc_ioc_swap_layouts(exp, karg);
2297 CERROR("unrecognised ioctl: cmd = %#x\n", cmd);
2298 GOTO(out, rc = -ENOTTY);
2301 module_put(THIS_MODULE);
2306 static int mdc_get_info_rpc(struct obd_export *exp,
2307 u32 keylen, void *key,
2308 u32 vallen, void *val)
2310 struct obd_import *imp = class_exp2cliimp(exp);
2311 struct ptlrpc_request *req;
2316 req = ptlrpc_request_alloc(imp, &RQF_MDS_GET_INFO);
2320 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_KEY,
2321 RCL_CLIENT, keylen);
2322 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_VALLEN,
2323 RCL_CLIENT, sizeof(vallen));
2325 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GET_INFO);
2327 ptlrpc_request_free(req);
2331 tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_KEY);
2332 memcpy(tmp, key, keylen);
2333 tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_VALLEN);
2334 memcpy(tmp, &vallen, sizeof(vallen));
2336 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_VAL,
2337 RCL_SERVER, vallen);
2338 ptlrpc_request_set_replen(req);
2340 /* if server failed to resolve FID, and OI scrub not able to fix it, it
2341 * will return -EINPROGRESS, ptlrpc_queue_wait() will keep retrying,
2342 * set request interruptible to avoid deadlock.
2344 if (KEY_IS(KEY_FID2PATH))
2345 req->rq_allow_intr = 1;
2347 rc = ptlrpc_queue_wait(req);
2348 /* -EREMOTE means the get_info result is partial, and it needs to
2349 * continue on another MDT, see fid2path part in lmv_iocontrol */
2350 if (rc == 0 || rc == -EREMOTE) {
2351 tmp = req_capsule_server_get(&req->rq_pill, &RMF_GETINFO_VAL);
2352 memcpy(val, tmp, vallen);
2353 if (req_capsule_rep_need_swab(&req->rq_pill)) {
2354 if (KEY_IS(KEY_FID2PATH))
2355 lustre_swab_fid2path(val);
2358 ptlrpc_req_finished(req);
2363 static void lustre_swab_hai(struct hsm_action_item *h)
2365 __swab32s(&h->hai_len);
2366 __swab32s(&h->hai_action);
2367 lustre_swab_lu_fid(&h->hai_fid);
2368 lustre_swab_lu_fid(&h->hai_dfid);
2369 __swab64s(&h->hai_cookie);
2370 __swab64s(&h->hai_extent.offset);
2371 __swab64s(&h->hai_extent.length);
2372 __swab64s(&h->hai_gid);
2375 static void lustre_swab_hal(struct hsm_action_list *h)
2377 struct hsm_action_item *hai;
2380 __swab32s(&h->hal_version);
2381 __swab32s(&h->hal_count);
2382 __swab32s(&h->hal_archive_id);
2383 __swab64s(&h->hal_flags);
2385 for (i = 0; i < h->hal_count; i++, hai = hai_next(hai))
2386 lustre_swab_hai(hai);
2389 static void lustre_swab_kuch(struct kuc_hdr *l)
2391 __swab16s(&l->kuc_magic);
2392 /* __u8 l->kuc_transport */
2393 __swab16s(&l->kuc_msgtype);
2394 __swab16s(&l->kuc_msglen);
2397 static int mdc_ioc_hsm_ct_start(struct obd_export *exp,
2398 struct lustre_kernelcomm *lk)
2400 struct obd_import *imp = class_exp2cliimp(exp);
2403 if (lk->lk_group != KUC_GRP_HSM) {
2404 CERROR("Bad copytool group %d\n", lk->lk_group);
2408 CDEBUG(D_HSM, "CT start r%d w%d u%d g%d f%#x\n", lk->lk_rfd, lk->lk_wfd,
2409 lk->lk_uid, lk->lk_group, lk->lk_flags);
2411 if (lk->lk_flags & LK_FLG_STOP) {
2412 /* Unregister with the coordinator */
2413 rc = mdc_ioc_hsm_ct_unregister(imp);
2415 __u32 *archives = NULL;
2417 if ((lk->lk_flags & LK_FLG_DATANR) && lk->lk_data_count > 0)
2418 archives = lk->lk_data;
2420 rc = mdc_ioc_hsm_ct_register(imp, lk->lk_data_count, archives);
2427 * Send a message to any listening copytools
2428 * @param val KUC message (kuc_hdr + hsm_action_list)
2429 * @param len total length of message
2431 static int mdc_hsm_copytool_send(const struct obd_uuid *uuid,
2432 size_t len, void *val)
2434 struct kuc_hdr *lh = (struct kuc_hdr *)val;
2435 struct hsm_action_list *hal = (struct hsm_action_list *)(lh + 1);
2439 if (len < sizeof(*lh) + sizeof(*hal)) {
2440 CERROR("Short HSM message %zu < %zu\n", len,
2441 sizeof(*lh) + sizeof(*hal));
2444 if (lh->kuc_magic == __swab16(KUC_MAGIC)) {
2445 lustre_swab_kuch(lh);
2446 lustre_swab_hal(hal);
2447 } else if (lh->kuc_magic != KUC_MAGIC) {
2448 CERROR("Bad magic %x!=%x\n", lh->kuc_magic, KUC_MAGIC);
2452 CDEBUG(D_HSM, " Received message mg=%x t=%d m=%d l=%d actions=%d "
2454 lh->kuc_magic, lh->kuc_transport, lh->kuc_msgtype,
2455 lh->kuc_msglen, hal->hal_count, hal->hal_fsname);
2457 /* Broadcast to HSM listeners */
2458 rc = libcfs_kkuc_group_put(uuid, KUC_GRP_HSM, lh);
2464 * callback function passed to kuc for re-registering each HSM copytool
2465 * running on MDC, after MDT shutdown/recovery.
2466 * @param data copytool registration data
2467 * @param cb_arg callback argument (obd_import)
2469 static int mdc_hsm_ct_reregister(void *data, void *cb_arg)
2471 struct obd_import *imp = (struct obd_import *)cb_arg;
2472 struct kkuc_ct_data *kcd = data;
2473 __u32 *archives = NULL;
2477 (kcd->kcd_magic != KKUC_CT_DATA_ARRAY_MAGIC &&
2478 kcd->kcd_magic != KKUC_CT_DATA_BITMAP_MAGIC))
2481 if (kcd->kcd_magic == KKUC_CT_DATA_BITMAP_MAGIC) {
2482 CDEBUG(D_HA, "%s: recover copytool registration to MDT "
2483 "(archive=%#x)\n", imp->imp_obd->obd_name,
2484 kcd->kcd_nr_archives);
2486 CDEBUG(D_HA, "%s: recover copytool registration to MDT "
2487 "(archive nr = %u)\n",
2488 imp->imp_obd->obd_name, kcd->kcd_nr_archives);
2489 if (kcd->kcd_nr_archives != 0)
2490 archives = kcd->kcd_archives;
2493 rc = mdc_ioc_hsm_ct_register(imp, kcd->kcd_nr_archives, archives);
2494 /* ignore error if the copytool is already registered */
2495 return (rc == -EEXIST) ? 0 : rc;
2499 * Re-establish all kuc contexts with MDT
2500 * after MDT shutdown/recovery.
2502 static int mdc_kuc_reregister(struct obd_import *imp)
2504 /* re-register HSM agents */
2505 return libcfs_kkuc_group_foreach(&imp->imp_obd->obd_uuid, KUC_GRP_HSM,
2506 mdc_hsm_ct_reregister, imp);
2509 static int mdc_set_info_async(const struct lu_env *env,
2510 struct obd_export *exp,
2511 u32 keylen, void *key,
2512 u32 vallen, void *val,
2513 struct ptlrpc_request_set *set)
2515 struct obd_import *imp = class_exp2cliimp(exp);
2519 if (KEY_IS(KEY_READ_ONLY)) {
2520 if (vallen != sizeof(int))
2523 spin_lock(&imp->imp_lock);
2524 if (*((int *)val)) {
2525 imp->imp_connect_flags_orig |= OBD_CONNECT_RDONLY;
2526 imp->imp_connect_data.ocd_connect_flags |=
2529 imp->imp_connect_flags_orig &= ~OBD_CONNECT_RDONLY;
2530 imp->imp_connect_data.ocd_connect_flags &=
2531 ~OBD_CONNECT_RDONLY;
2533 spin_unlock(&imp->imp_lock);
2535 rc = do_set_info_async(imp, MDS_SET_INFO, LUSTRE_MDS_VERSION,
2536 keylen, key, vallen, val, set);
2539 if (KEY_IS(KEY_CHANGELOG_CLEAR)) {
2540 rc = do_set_info_async(imp, MDS_SET_INFO, LUSTRE_MDS_VERSION,
2541 keylen, key, vallen, val, set);
2544 if (KEY_IS(KEY_HSM_COPYTOOL_SEND)) {
2545 rc = mdc_hsm_copytool_send(&imp->imp_obd->obd_uuid, vallen,
2550 if (KEY_IS(KEY_DEFAULT_EASIZE)) {
2551 __u32 *default_easize = val;
2553 exp->exp_obd->u.cli.cl_default_mds_easize = *default_easize;
2557 rc = osc_set_info_async(env, exp, keylen, key, vallen, val, set);
2561 static int mdc_get_info(const struct lu_env *env, struct obd_export *exp,
2562 __u32 keylen, void *key, __u32 *vallen, void *val)
2566 if (KEY_IS(KEY_MAX_EASIZE)) {
2567 __u32 mdsize, *max_easize;
2569 if (*vallen != sizeof(int))
2571 mdsize = *(__u32 *)val;
2572 if (mdsize > exp->exp_obd->u.cli.cl_max_mds_easize)
2573 exp->exp_obd->u.cli.cl_max_mds_easize = mdsize;
2575 *max_easize = exp->exp_obd->u.cli.cl_max_mds_easize;
2577 } else if (KEY_IS(KEY_DEFAULT_EASIZE)) {
2578 __u32 *default_easize;
2580 if (*vallen != sizeof(int))
2582 default_easize = val;
2583 *default_easize = exp->exp_obd->u.cli.cl_default_mds_easize;
2585 } else if (KEY_IS(KEY_CONN_DATA)) {
2586 struct obd_import *imp = class_exp2cliimp(exp);
2587 struct obd_connect_data *data = val;
2589 if (*vallen != sizeof(*data))
2592 *data = imp->imp_connect_data;
2594 } else if (KEY_IS(KEY_TGT_COUNT)) {
2595 *((__u32 *)val) = 1;
2599 rc = mdc_get_info_rpc(exp, keylen, key, *vallen, val);
2604 static int mdc_fsync(struct obd_export *exp, const struct lu_fid *fid,
2605 struct ptlrpc_request **request)
2607 struct ptlrpc_request *req;
2612 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_SYNC);
2616 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_SYNC);
2618 ptlrpc_request_free(req);
2622 mdc_pack_body(&req->rq_pill, fid, 0, 0, -1, 0);
2624 ptlrpc_request_set_replen(req);
2626 rc = ptlrpc_queue_wait(req);
2628 ptlrpc_req_finished(req);
2634 struct mdc_rmfid_args {
2639 int mdc_rmfid_interpret(const struct lu_env *env, struct ptlrpc_request *req,
2642 struct mdc_rmfid_args *aa;
2647 aa = ptlrpc_req_async_args(aa, req);
2649 size = req_capsule_get_size(&req->rq_pill, &RMF_RCS,
2651 LASSERT(size == sizeof(int) * aa->mra_nr);
2652 rcs = req_capsule_server_get(&req->rq_pill, &RMF_RCS);
2654 LASSERT(aa->mra_rcs);
2655 LASSERT(aa->mra_nr);
2656 memcpy(aa->mra_rcs, rcs, size);
2662 static int mdc_rmfid(struct obd_export *exp, struct fid_array *fa,
2663 int *rcs, struct ptlrpc_request_set *set)
2665 struct ptlrpc_request *req;
2666 struct mdc_rmfid_args *aa;
2672 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_RMFID);
2676 flen = fa->fa_nr * sizeof(struct lu_fid);
2677 req_capsule_set_size(&req->rq_pill, &RMF_FID_ARRAY,
2679 req_capsule_set_size(&req->rq_pill, &RMF_FID_ARRAY,
2681 req_capsule_set_size(&req->rq_pill, &RMF_RCS,
2682 RCL_SERVER, fa->fa_nr * sizeof(__u32));
2683 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_RMFID);
2685 ptlrpc_request_free(req);
2688 tmp = req_capsule_client_get(&req->rq_pill, &RMF_FID_ARRAY);
2689 memcpy(tmp, fa->fa_fids, flen);
2691 mdc_pack_body(&req->rq_pill, NULL, 0, 0, -1, 0);
2692 b = req_capsule_client_get(&req->rq_pill, &RMF_MDT_BODY);
2693 b->mbo_ctime = ktime_get_real_seconds();
2695 ptlrpc_request_set_replen(req);
2698 aa = ptlrpc_req_async_args(aa, req);
2700 aa->mra_nr = fa->fa_nr;
2701 req->rq_interpret_reply = mdc_rmfid_interpret;
2703 ptlrpc_set_add_req(set, req);
2704 ptlrpc_check_set(NULL, set);
2709 static int mdc_import_event(struct obd_device *obd, struct obd_import *imp,
2710 enum obd_import_event event)
2712 struct client_obd *cli = &obd->u.cli;
2715 LASSERT(imp->imp_obd == obd);
2718 case IMP_EVENT_DISCON:
2719 spin_lock(&cli->cl_loi_list_lock);
2720 cli->cl_avail_grant = 0;
2721 cli->cl_lost_grant = 0;
2722 spin_unlock(&cli->cl_loi_list_lock);
2724 case IMP_EVENT_INACTIVE:
2726 * Flush current sequence to make client obtain new one
2727 * from server in case of disconnect/reconnect.
2729 down_read(&cli->cl_seq_rwsem);
2731 seq_client_flush(cli->cl_seq);
2732 up_read(&cli->cl_seq_rwsem);
2734 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE);
2736 case IMP_EVENT_INVALIDATE: {
2737 struct ldlm_namespace *ns = obd->obd_namespace;
2741 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
2743 env = cl_env_get(&refcheck);
2745 /* Reset grants. All pages go to failing rpcs due to
2746 * the invalid import.
2748 osc_io_unplug(env, cli, NULL);
2750 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2751 osc_ldlm_resource_invalidate,
2753 cl_env_put(env, &refcheck);
2754 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
2760 case IMP_EVENT_ACTIVE:
2761 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE);
2762 /* redo the kuc registration after reconnecting */
2764 rc = mdc_kuc_reregister(imp);
2766 case IMP_EVENT_OCD: {
2767 struct obd_connect_data *ocd = &imp->imp_connect_data;
2769 if (OCD_HAS_FLAG(ocd, GRANT))
2770 osc_init_grant(cli, ocd);
2772 md_init_ea_size(obd->obd_self_export, ocd->ocd_max_easize, 0);
2773 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD);
2776 case IMP_EVENT_DEACTIVATE:
2777 case IMP_EVENT_ACTIVATE:
2780 CERROR("Unknown import event %x\n", event);
2786 int mdc_fid_alloc(const struct lu_env *env, struct obd_export *exp,
2787 struct lu_fid *fid, struct md_op_data *op_data)
2789 struct client_obd *cli = &exp->exp_obd->u.cli;
2794 down_read(&cli->cl_seq_rwsem);
2796 rc = seq_client_alloc_fid(env, cli->cl_seq, fid);
2797 up_read(&cli->cl_seq_rwsem);
2802 static struct obd_uuid *mdc_get_uuid(struct obd_export *exp)
2804 struct client_obd *cli = &exp->exp_obd->u.cli;
2805 return &cli->cl_target_uuid;
2809 * Determine whether the lock can be canceled before replaying it during
2810 * recovery, non zero value will be return if the lock can be canceled,
2811 * or zero returned for not
2813 static int mdc_cancel_weight(struct ldlm_lock *lock)
2815 if (lock->l_resource->lr_type != LDLM_IBITS)
2818 /* FIXME: if we ever get into a situation where there are too many
2819 * opened files with open locks on a single node, then we really
2820 * should replay these open locks to reget it */
2821 if (lock->l_policy_data.l_inodebits.bits & MDS_INODELOCK_OPEN)
2824 /* Special case for DoM locks, cancel only unused and granted locks */
2825 if (ldlm_has_dom(lock) &&
2826 (lock->l_granted_mode != lock->l_req_mode ||
2827 osc_ldlm_weigh_ast(lock) != 0))
2833 static int mdc_resource_inode_free(struct ldlm_resource *res)
2835 if (res->lr_lvb_inode)
2836 res->lr_lvb_inode = NULL;
2841 static struct ldlm_valblock_ops inode_lvbo = {
2842 .lvbo_free = mdc_resource_inode_free
2845 static int mdc_llog_init(struct obd_device *obd)
2847 struct obd_llog_group *olg = &obd->obd_olg;
2848 struct llog_ctxt *ctxt;
2853 rc = llog_setup(NULL, obd, olg, LLOG_CHANGELOG_REPL_CTXT, obd,
2858 ctxt = llog_group_get_ctxt(olg, LLOG_CHANGELOG_REPL_CTXT);
2859 llog_initiator_connect(ctxt);
2860 llog_ctxt_put(ctxt);
2865 static void mdc_llog_finish(struct obd_device *obd)
2867 struct llog_ctxt *ctxt;
2871 ctxt = llog_get_context(obd, LLOG_CHANGELOG_REPL_CTXT);
2873 llog_cleanup(NULL, ctxt);
2878 int mdc_setup(struct obd_device *obd, struct lustre_cfg *cfg)
2884 rc = osc_setup_common(obd, cfg);
2888 rc = mdc_tunables_init(obd);
2890 GOTO(err_osc_cleanup, rc);
2892 obd->u.cli.cl_dom_min_inline_repsize = MDC_DOM_DEF_INLINE_REPSIZE;
2893 obd->u.cli.cl_lsom_update = true;
2895 ns_register_cancel(obd->obd_namespace, mdc_cancel_weight);
2897 obd->obd_namespace->ns_lvbo = &inode_lvbo;
2899 rc = mdc_llog_init(obd);
2901 CERROR("%s: failed to setup llogging subsystems: rc = %d\n",
2903 GOTO(err_llog_cleanup, rc);
2906 rc = mdc_changelog_cdev_init(obd);
2908 CERROR("%s: failed to setup changelog char device: rc = %d\n",
2910 GOTO(err_changelog_cleanup, rc);
2915 err_changelog_cleanup:
2916 mdc_llog_finish(obd);
2918 lprocfs_free_md_stats(obd);
2919 ptlrpc_lprocfs_unregister_obd(obd);
2921 osc_cleanup_common(obd);
2925 /* Initialize the default and maximum LOV EA sizes. This allows
2926 * us to make MDS RPCs with large enough reply buffers to hold a default
2927 * sized EA without having to calculate this (via a call into the
2928 * LOV + OSCs) each time we make an RPC. The maximum size is also tracked
2929 * but not used to avoid wastefully vmalloc()'ing large reply buffers when
2930 * a large number of stripes is possible. If a larger reply buffer is
2931 * required it will be reallocated in the ptlrpc layer due to overflow.
2933 static int mdc_init_ea_size(struct obd_export *exp, __u32 easize,
2936 struct obd_device *obd = exp->exp_obd;
2937 struct client_obd *cli = &obd->u.cli;
2940 if (cli->cl_max_mds_easize < easize)
2941 cli->cl_max_mds_easize = easize;
2943 if (cli->cl_default_mds_easize < def_easize)
2944 cli->cl_default_mds_easize = def_easize;
2949 static int mdc_precleanup(struct obd_device *obd)
2953 osc_precleanup_common(obd);
2955 mdc_changelog_cdev_finish(obd);
2956 mdc_llog_finish(obd);
2957 lprocfs_free_md_stats(obd);
2958 ptlrpc_lprocfs_unregister_obd(obd);
2963 static int mdc_cleanup(struct obd_device *obd)
2965 return osc_cleanup_common(obd);
2968 static const struct obd_ops mdc_obd_ops = {
2969 .o_owner = THIS_MODULE,
2970 .o_setup = mdc_setup,
2971 .o_precleanup = mdc_precleanup,
2972 .o_cleanup = mdc_cleanup,
2973 .o_add_conn = client_import_add_conn,
2974 .o_del_conn = client_import_del_conn,
2975 .o_connect = client_connect_import,
2976 .o_reconnect = osc_reconnect,
2977 .o_disconnect = osc_disconnect,
2978 .o_iocontrol = mdc_iocontrol,
2979 .o_set_info_async = mdc_set_info_async,
2980 .o_statfs = mdc_statfs,
2981 .o_statfs_async = mdc_statfs_async,
2982 .o_fid_init = client_fid_init,
2983 .o_fid_fini = client_fid_fini,
2984 .o_fid_alloc = mdc_fid_alloc,
2985 .o_import_event = mdc_import_event,
2986 .o_get_info = mdc_get_info,
2987 .o_get_uuid = mdc_get_uuid,
2988 .o_quotactl = mdc_quotactl,
2991 static const struct md_ops mdc_md_ops = {
2992 .m_get_root = mdc_get_root,
2993 .m_null_inode = mdc_null_inode,
2994 .m_close = mdc_close,
2995 .m_create = mdc_create,
2996 .m_enqueue = mdc_enqueue,
2997 .m_getattr = mdc_getattr,
2998 .m_getattr_name = mdc_getattr_name,
2999 .m_intent_lock = mdc_intent_lock,
3001 .m_rename = mdc_rename,
3002 .m_setattr = mdc_setattr,
3003 .m_setxattr = mdc_setxattr,
3004 .m_getxattr = mdc_getxattr,
3005 .m_fsync = mdc_fsync,
3006 .m_file_resync = mdc_file_resync,
3007 .m_read_page = mdc_read_page,
3008 .m_unlink = mdc_unlink,
3009 .m_cancel_unused = mdc_cancel_unused,
3010 .m_init_ea_size = mdc_init_ea_size,
3011 .m_set_lock_data = mdc_set_lock_data,
3012 .m_lock_match = mdc_lock_match,
3013 .m_get_lustre_md = mdc_get_lustre_md,
3014 .m_free_lustre_md = mdc_free_lustre_md,
3015 .m_set_open_replay_data = mdc_set_open_replay_data,
3016 .m_clear_open_replay_data = mdc_clear_open_replay_data,
3017 .m_intent_getattr_async = mdc_intent_getattr_async,
3018 .m_revalidate_lock = mdc_revalidate_lock,
3019 .m_rmfid = mdc_rmfid,
3022 dev_t mdc_changelog_dev;
3023 struct class *mdc_changelog_class;
3024 static int __init mdc_init(void)
3027 rc = alloc_chrdev_region(&mdc_changelog_dev, 0,
3028 MDC_CHANGELOG_DEV_COUNT,
3029 MDC_CHANGELOG_DEV_NAME);
3033 mdc_changelog_class = class_create(THIS_MODULE, MDC_CHANGELOG_DEV_NAME);
3034 if (IS_ERR(mdc_changelog_class)) {
3035 rc = PTR_ERR(mdc_changelog_class);
3039 rc = class_register_type(&mdc_obd_ops, &mdc_md_ops, true,
3040 LUSTRE_MDC_NAME, &mdc_device_type);
3047 class_destroy(mdc_changelog_class);
3049 unregister_chrdev_region(mdc_changelog_dev, MDC_CHANGELOG_DEV_COUNT);
3053 static void __exit mdc_exit(void)
3055 class_unregister_type(LUSTRE_MDC_NAME);
3056 class_destroy(mdc_changelog_class);
3057 unregister_chrdev_region(mdc_changelog_dev, MDC_CHANGELOG_DEV_COUNT);
3058 idr_destroy(&mdc_changelog_minor_idr);
3061 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3062 MODULE_DESCRIPTION("Lustre Metadata Client");
3063 MODULE_VERSION(LUSTRE_VERSION_STRING);
3064 MODULE_LICENSE("GPL");
3066 module_init(mdc_init);
3067 module_exit(mdc_exit);